Efficiency is now the headline. On Aug 24, 2026, NVIDIA’s latest system figures placed Vera Rubin and Blackwell at the center of a performance-per-watt contest for agentic AI inference.
The key measure is AI-factory throughput per megawatt, not a standalone speed claim. On AgentX workloads, NVIDIA Vera Rubin NVL72 achieved up to 30x higher throughput per megawatt than GB300 NVL72, turning the comparison into a direct test of how much inference output each system can produce from the same power budget.
That gap matters because agentic AI workloads can involve long sequences of model inference rather than one isolated response. The verified result does not claim that Vera Rubin is 30x faster in every task; it reports up to 30x higher AI-factory throughput per megawatt on AgentX workloads. Precision matters, especially when the number is large enough to attract its own marketing department.
Vera Rubin widens the efficiency gap
NVIDIA Vera Rubin NVL72 sits at the top of the first comparison, with GB300 NVL72 as the reference point. The result gives Vera Rubin a clear advantage in the stated AgentX test, but the measurement remains tied to throughput per megawatt rather than raw throughput alone.
That distinction changes how the systems should be read. A higher throughput-per-megawatt result means the system delivers more AI-factory throughput within a given power measure; it does not, by itself, provide a complete ranking for every model, workload, or deployment.
The comparison also shows that NVIDIA is presenting system-level efficiency as a central part of its agentic AI strategy. The named systems are NVL72 configurations, and the claim concerns the complete AI-factory throughput metric rather than a single component viewed in isolation.
Blackwell pushes its own generational gains
GB300 NVL72 has another result attached to it: up to 80x gains in throughput per megawatt over H200 NVL8 for large MoE models such as Kimi K3 2.8T. That is a separate comparison from the 30x Vera Rubin result, with a different baseline, configuration, and model category.
The distinction is important because the two numbers answer different questions. The 30x figure compares NVIDIA Vera Rubin NVL72 with GB300 NVL72 on AgentX workloads, while the 80x figure describes GB300 NVL72’s multi-generational advantage over H200 NVL8 for large MoE models such as Kimi K3 2.8T.
Put together, the figures describe two layers of progress. Vera Rubin reaches a higher stated efficiency result against GB300 NVL72 on AgentX, while GB300 NVL72 extends its own throughput-per-megawatt advantage over H200 NVL8 in a large-model comparison.
Neither number should be flattened into a universal performance claim. The workloads and comparison systems differ, and both results use the word “up to,” which marks the strongest reported outcome rather than a promise that applies to every run.
Still, the direction is clear within the supplied comparisons: NVIDIA is measuring agentic AI progress through the amount of inference work completed for each megawatt. For systems built to run large models and AI agents, that metric puts hardware capability and power use in the same frame.
The Vera Rubin NVL72 and GB300 NVL72 results therefore matter less as isolated bragging rights than as evidence of a changing benchmark. AgentX workloads, large MoE models, and configurations such as H200 NVL8 are being compared through efficiency as much as speed. The hardware race has acquired a power bill.
